philbert440/Qwen3.6-27B-Uncensored-Aggressive
philbert440/Qwen3.6-27B-Uncensored-Aggressive is a 27 billion parameter vision-language model based on Qwen3.6-27B, developed by Rivet. This model has been processed with Heretic (KL-optimized abliteration) to significantly reduce refusal behaviors across all categories, including sensitive topics. It preserves vision, video, and MTP speculative-decoding capabilities, offering maximum openness for research, red-teaming, and unrestricted assistant use. The model maintains coherence on neutral long-form prompts while providing real, substantive content in response to previously refused queries.
Loading preview...
Model Overview
philbert440/Qwen3.6-27B-Uncensored-Aggressive is a 27 billion parameter vision-language model derived from Qwen3.6-27B. Developed by Rivet, this model has undergone a specialized process called Heretic (KL-optimized, per-layer continuous abliteration) to eliminate refusal behaviors. It is designed for maximum openness, making it suitable for research, red-teaming, and unrestricted assistant applications.
Key Capabilities & Differentiators
- Refusal Suppression: Significantly reduces refusals across a wide range of categories, including coding/hacking, malware, lockpicking, disinformation, weapons, drugs, and explicit content.
- Content Generation: Provides real and substantive content for requests that the base model would typically refuse, such as functional code examples or conceptual explanations for sensitive topics.
- Vision & MTP Preservation: Fully retains the vision, video, and MTP speculative-decoding head from the base Qwen3.6-27B model.
- Coherence: Maintains high coherence and provides full-length, on-topic answers on neutral long-form prompts.
- Aggressive Tier: Represents an "Aggressive" tier of openness, accepting a larger shift from the base distribution to prioritize unrestricted output.
Method & Limitations
The model was created using Heretic's multi-objective optimization, co-minimizing refusals and KL-divergence from the base model. This involved per-layer continuous ablation of residual-writing matrices. While it provides correct conceptual content for topics like weapons and drugs, it does not hallucinate actionable recipe-level detail if that knowledge is not present in the base weights. As an "Aggressive" variant, it trades some fidelity for openness, resulting in perplexity shifts compared to the base model.